{"slug": "openai-is-bringing-in-the-hackers-who-got-inside-anthropic-to-test-its-own", "title": "OpenAI is bringing in the hackers who got inside Anthropic to test its own defences", "summary": "OpenAI said on Monday it will bring in security start-up Cathedral, whose AI agents reached parts of Anthropic's internal systems in under 12 hours last month, to attack its own defenses during its next internal security sprint, and it is more than quadrupling the midpoint of its bug bounty reward ranges for medium, high and critical findings, offering up to $172,000 for the most damaging finds. Cathedral, founded by four former DOGE staffers, found routes to Anthropic's Slack messages, meeting transcripts and salary records through flaws in Anthropic's business software rather than Claude itself, per The Washington Post's October 26 report; Anthropic said Cathedral reached only \"a limited set\" of its sales, HR and IT systems and that it cut off access the same day. OpenAI's prior sprint put 25% of its product engineers on finding and fixing weaknesses, mobilizing more than 250 people, and the company repeated its pledge of $1 billion in subsidized access over six months to its most capable cyber models through its Daybreak programme.", "body_md": "The OpenAI logo. Image: OpenAI logo via LobeHub Icons (MIT); trademark of OpenAI\n\nOpenAI is bringing in Cathedral, the security start-up whose AI agents got inside parts of Anthropic’s internal systems last month, to attack its own defences. OpenAI [said on Monday](https://openai.com/index/preparing-our-next-security-sprint-to-harden-openai/) that Cathedral’s researchers will use OpenAI’s models alongside its own teams in its next internal security sprint, to “challenge our defenses”.\n\nAt the same time it is raising what it pays outside researchers for finding holes in its products, more than quadrupling typical rewards for serious bugs and offering up to $172,000 for the most damaging finds.\n\n## From Anthropic’s systems to OpenAI’s\n\nCathedral, founded by four former DOGE staffers, says its AI agents [found routes to Anthropic’s Slack messages](https://madrobot.blog/2026/10/26/ex-doge-cathedral-ai-agents-inside-anthropic-12-hours/), meeting transcripts and salary records in under 12 hours, [The Washington Post reported](https://www.washingtonpost.com/technology/2026/10/26/researchers-cybersecurity-start-up-cathedral-got-inside-anthropic/) on October 26. It got in through flaws in business software Anthropic uses for sales and admin work, not through Claude itself, and reported what it found through Anthropic’s bug bounty programme.\n\nAnthropic’s account was narrower. It said Cathedral reached only “a limited set” of its sales, HR and IT systems and that it shut off the start-up’s access the same day. It also accused Cathedral of overstating the national security stakes and of not fully following the bounty rules, because it did not disclose all of its activity.\n\nOpenAI’s announcement doesn’t mention any of that. It says only that Cathedral will bring “differentiated expertise and approaches to finding weaknesses”. Neither company has said when the sprint starts, how long it will run or what Cathedral will be paid.\n\n## A quarter of OpenAI’s product engineers\n\nThis will be OpenAI’s second sprint of this kind. In the last one, it says, it put 25% of its product engineers onto finding and fixing weaknesses in its systems, using its Codex coding agent to investigate findings, reproduce vulnerabilities in isolated environments and prepare patches for review. A [write-up of that work](https://openai.com/the-defense-factory/), which OpenAI calls its Defense Factory, says it mobilised more than 250 people and treated the job “with the urgency of an incident response”.\n\nOpenAI’s case is that “traditional cyber defenses alone are no longer sufficient”, citing “recent incidents across the industry”. It argues defenders have a head start, which it calls the defender’s window, because they can point the most capable models at their own code before attackers get similar power from [open-weight models anyone can download](https://madrobot.blog/2026/09/29/anthropic-glm-5-3-zai-cyber-exploits-safeguards-open-weight/). Its models are already finding real bugs: in October one helped uncover [two critical flaws in PHP](https://madrobot.blog/2026/10/21/php2shell-critical-php-flaws-openai-gpt-5-6-sol-calif-wiz-patch-now/).\n\nOpenAI also repeated its pledge of $1 billion in subsidised access over six months to its most capable cyber models for “frontline defenders protecting essential services and digital infrastructure”, through its Daybreak programme. Ukraine’s government is [getting Daybreak for free](https://madrobot.blog/2026/09/23/openai-daybreak-cyber-defence-ukraine-free/), and Google has taken a similar line, releasing Gemini 4 Carbon [to cyber defenders first](https://madrobot.blog/2026/10/28/google-gemini-4-carbon-tops-leaderboard-cyber-defenders/).\n\n## Bug bounties of up to $172,000\n\nThe [changes to OpenAI’s Security Bug Bounty](https://openai.com/index/rewarding-defenders-with-bigger-bug-bounties/), which runs on [Bugcrowd](https://bugcrowd.com/engagements/openai), take effect now:\n\n- **Higher base rewards.** OpenAI says it is more than quadrupling the midpoint of its reward ranges for medium, high and critical findings, on average. A critical bug in ChatGPT, Codex or the API Platform now earns up to $30,000.\n- **Bonus multipliers.** Reports with a reliable proof of concept get a quality bonus, a new five-tier loyalty scheme rewards repeat researchers, and short “priority campaigns” add bonus pools for newly launched features.\n- **A higher ceiling.** Exceptional critical findings in those core products can still earn up to $100,000 before bonuses, and up to $172,000 with them.\n- **Back pay.** Reports still waiting for a reward get the new rates, and anyone paid in the 30 days before the announcement gets the difference if the new ranges would have paid more.\n\nOpenAI’s own examples, which it calls illustrative, show the size of the jump. A one-click ChatGPT account takeover that once earned $6,500 would now pay $51,577 with the top bonuses stacked, and a flaw letting someone read any ChatGPT user’s chats would pay $171,925.\n\nIts security team will also be at the PROMPT2PWN event in Oakland, California, on November 20 and 21, where researchers can test targets in person.\n\n## Why it matters\n\nA week after Anthropic accused Cathedral of breaking its bug bounty rules, its biggest rival has hired the start-up to attack its own systems. Together with much bigger bounties, it shows how seriously AI labs now take the threat of AI-powered attackers, and how much they are willing to pay to find their own weak spots first.", "url": "https://wpnews.pro/news/openai-is-bringing-in-the-hackers-who-got-inside-anthropic-to-test-its-own", "canonical_source": "https://madrobot.blog/2026/11/03/openai-cathedral-security-sprint-bigger-bug-bounties/", "published_at": "2026-11-03 08:32:00+00:00", "updated_at": "2026-11-03 09:07:15.874662+00:00", "lang": "en", "topics": ["ai-safety", "artificial-intelligence", "ai-agents"], "entities": ["OpenAI", "Cathedral", "Anthropic", "Claude", "Bugcrowd", "Daybreak", "The Washington Post", "Codex"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/openai-is-bringing-in-the-hackers-who-got-inside-anthropic-to-test-its-own", "markdown": "https://wpnews.pro/news/openai-is-bringing-in-the-hackers-who-got-inside-anthropic-to-test-its-own.md", "text": "https://wpnews.pro/news/openai-is-bringing-in-the-hackers-who-got-inside-anthropic-to-test-its-own.txt", "jsonld": "https://wpnews.pro/news/openai-is-bringing-in-the-hackers-who-got-inside-anthropic-to-test-its-own.jsonld"}}